AI infrastructurein a glass boxyou own
Build and ship AI applications on infrastructure you own — your hosts, your keys, your choice of model.
- // last 30 days on main
- 925 commits
- 525 PRs merged
- 565 issues closed
The platform, in four numbers.
No asterisks — every figure is countable in the repo. Nine modules ship today and two are still on the roadmap; the manifest below names which.
Nine modules. One toolkit.
The nine that ship run standalone or compose with the rest — retrieval, quant research and chat, plus the auth, tracing and audit any deployment needs. Two more are on the roadmap; the manifest below marks which.
select a module
Built on what you already run.
Every module is assembled from open-source libraries you can name, version, and swap — no forks, no reimplementations. That is why this drops into an existing deployment instead of asking you to replace one.
Maintained in the open.
One MIT-licensed monorepo, public and moving — 12 issues open today against 565 closed in the last 30 days. Below are the most recent features to land on main — each one a pull request you can open and read, not a changelog entry we wrote about ourselves. Figures are a snapshot taken 2026-08-17.
- #2369rebalance model pools for cost and financial-analysis capabilityolympus · 2026-08-15
- #2357BYOK provider/model catalog + live model list at key stepdigichat · 2026-08-14
- #2361add x.ai to the routable-provider allowlist (digigraph)digigraph · 2026-08-13
- #2346Store API for cross-thread, per-subject preference memorydigigraph · 2026-08-13
- #2322emit a round_boundary signal when a round narrates alongside tool callsdigigraph,digillm · 2026-08-13
- #2324bring mermaid/code-block rendering to the primary chat pagedigichat,digiweb · 2026-08-13
git clone https://github.com/digithings-ai/latest release digichat-v1.3.0 · 2026-08-15
Issues are open and the branching model is written down. Pick one up, or open an issue for the module you need.
Nobody knows which vendor wins.
The field moves faster than any bet you could place on it, so this architecture declines to place one. Nothing here is married to a provider: the stack is self-hosted, keys arrive per request, and models and vector backends are configuration rather than code. When the leader changes, you change one setting — not the application on top.
Self-hosted by default
One docker-compose file, sixteen services, on a laptop or any host you own — eight up by default and the rest behind named profiles you turn on: digichat, digivault, the search MCP, the heartbeat, observability, and the LiteLLM cache.
BYOK, every request
Anthropic, OpenAI, or any LiteLLM-compatible key — forwarded per-request, never stored.
Audit-on by default
Append-only JSONL audit and a correlation ID on every service hop. Events record a prompt's length, never its text — tail events.jsonl and check.
Backend-swappable
Two live vector backends — Chroma and Azure AI Search — behind one client, swappable without touching business code.
Questions, enterprise, or partnership.
The whole monorepo is MIT-licensed and public — take it and run it yourself. What we sell is the integration work: fitting these modules to the stack you already have, on your own infrastructure.